On this page

Synthesis: Akkurt, Maurya and Brown (2025) asked counselor trainees to practice with GPT-4o prompted to behave as a client, then coded the session transcripts and ran a focus group with the four trainees who completed the study. The simulator gave them something a peer role-play rarely does: a psychologically safe place to be clumsy, which lowered performance anxiety and built confidence before the first real client. The limits were equally clear and all on the realism side — a client who agrees too readily, no emotional nuance in the text, and cultural neutrality that erases exactly the dynamics counseling training works hardest to develop.

Key Findings

  • Safety, not realism, is what the simulator supplied. Trainees described a flexible, low-stakes environment where a mistake cost nothing, which reduced performance anxiety and raised confidence about working with real clients.
  • Prompt detail governs quality. Participants said the simulation's realism and complexity tracked how much context they put into the prompt, which makes prompt construction part of the counseling preparation rather than a technical afterthought.
  • Agreeableness is the main artifact. The simulated client was overly agreeable, which removes the resistance and ambivalence trainees most need to practice working with.
  • Emotional and cultural flatness limits transfer. The transcripts showed little emotional nuance and a cultural neutrality that skipped the identity dynamics counseling competence depends on.
  • Uptake was low even in a recruited sample. 52 students consented, nine submitted transcripts, and four joined the focus group out of 212 invited.

What the study did

The design is a qualitative content analysis. Trainees were given a prompt specification, ran a 30–40 minute role-play with GPT-4o as client, and submitted the transcript; within two to four weeks they received written feedback and were invited to a focus group. Transcripts were coded with a concept-driven framework covering authenticity, emotional expression, consistency, self-awareness and cultural dynamics, and the focus-group data were coded inductively.

The nine transcripts are the study's real evidence, and the coding is consensus-based between the first two authors rather than independent with reliability statistics. The focus group adds the trainees' own account of what the exercise did for them, which is where the anxiety and confidence findings come from.

Why this belongs in the counseling curriculum as a rehearsal, not a substitute

The distinction the trainees drew is the useful one. Practicing with a simulator before a live client is rehearsal: it lowers the cost of the first attempts and lets a trainee hear their own wording. It is not a substitute for a client who resists, because the simulator's agreeableness removes the clinical work. A trainee who only ever practises against a compliant client learns to explain, not to work with ambivalence.

The cultural finding is sharper. A model that returns culturally neutral responses does not fail to notice culture — it actively normalises not noticing it, in the one setting where noticing it is the competence being assessed. Any program adopting this needs to pair the simulation with human role-play that carries the identity dynamics, and to say so in the assignment rather than leaving it to chance.

What this means for practice

  • Instructors: use the simulator for first-attempt rehearsal and anxiety reduction, and schedule human role-play for resistance, silence and cultural dynamics.
  • Instructors: make the prompt itself assessable. Trainees who wrote detailed context got a more useful client, and prompt writing is a form of case conceptualisation.
  • Clinical educators: build in a debrief on what the simulator got wrong. The agreeableness and cultural flatness are teaching material, not defects to hide.
  • Programmes: do not read low voluntary uptake as low interest. Nine of 212 invited students submitted transcripts, which says something about workload and incentives.

Limitations

Nine transcripts and four focus-group participants, from one counseling program, with consensus coding and no reliability statistics. The study reports trainee perceptions and coded transcript features rather than client outcomes or transfer to real sessions. Prompt quality varied across participants, so the "client" was not the same stimulus for everyone. Published in 2025, it describes GPT-4o behavior that has since changed.

Citation

Akkurt, M., Maurya, R., & Brown, T. (2025). Learning Through Simulation: Counselor Trainees' Interactions with ChatGPT as a Client. Behavioral Sciences, 15(1660).

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.